伪异常引导的融合注意力和记忆增强的鱼群异常行为检测  

Pseudo-anomaly-guided fish school abnormal behavior detection with fused attention and memory augmentation

作　　者：袁红春[1] 肖智豪 YUAN Hongchun;XIAO Zhihao(School of Information,Shanghai Ocean University,Shanghai 201306,China)

机构地区：[1]上海海洋大学信息学院,上海201306

出　　处：《湖南农业大学学报(自然科学版)》2025年第1期123-130,共8页Journal of Hunan Agricultural University(Natural Sciences)

基　　金：国家自然科学基金项目(32403031)。

摘　　要：现有的鱼群异常行为检测方法无法有效提取高级语义信息、特征学习不足,且缺乏对异常样本的学习和提取关键特征的能力,无法满足现有的大规模水产养殖需求。笔者结合深度学习技术,提出了一种伪异常引导的融合注意力和记忆增强的鱼群异常行为检测方法:通过在视频序列中随机选择跳跃的帧构建伪异常合成器生成伪异常样本,增强对异常样本的感知能力;提出选择性内核频率通道注意力(SKFca)机制,在选择性内核(SK)注意力的基础上引入频域信息,以捕捉更丰富的输入信息;通过瓶颈注意力(BAM)机制在通道和空间维度上抑制不相关的背景特征,突出前景目标特征;在2种注意力模块后面添加记忆增强模块,将异常样本的编码特征替换为正常样本的编码特征,扩大异常样本输出与输入的重构误差;将记忆增强后的通道和空间维度上的关键特征和频域特征融合,以全面提取高级语义信息。结果表明,本研究所提方法在2种自制的鱼类数据集上检测效果都很好,曲线下面积(AUC)分别达0.953和0.957,且能实现对异常的精确定位。Existing methods for detecting abnormal behavior in fish schools had difficulties to extract higher-level semantic information and learn features effectively,or identify key features of anomalies,making them unsuitable for large-scale aquaculture.To address this,we proposed a deep learning-based fish school abnormal behavior detection method combining pseudo-anomaly guidance,fused attention,and memory augmentation.First,a pseudo-anomaly synthesizer was developed to enhance anomaly perception by randomly skipping frames in video sequences to generate pseudo-anomalous samples.Next,the SKFca attention mechanism integrated frequency domain information into the SK attention mechanism to capture richer input features,while the BAM attention mechanism suppressed irrelevant background features in channel and spatial dimensions to emphasize foreground targets.A memory-augmented module replaced with encoded anomaly features with normal samples,amplifying reconstruction errors for anomalies.Finally,memory-augmented key features and enriched frequency domain features were merged to extract comprehensive high-level semantic information.Experiments results on two self-made fish datasets demonstrated superior performance,achieving AUC values of 0.953 and 0.957 with precise anomaly localization.

关 键 词：鱼群异常行为检测 高级语义信息 深度学习 伪异常引导 注意力机制 记忆增强 

分 类 号：S951.2[农业科学—水产养殖] TP391.41[农业科学—水产科学]